2.1

Feature extraction of insect image

After the original image is preprocessed to get the segmented image, we also need to describe the attributes of the object so that the computer can understand the segmented image and then distinguish and classify it. The attributes of an object are called features, and the process of describing an object is called feature extraction. Feature extraction is the most critical link in image recognition, which strongly affects the design and performance of classifier. In traditional insect taxonomy, the body structure of insects, such as body length and size, is an important basis for identification and classification, and these characteristics generally belong to the description of shape characteristics.

Generally speaking, when computer vision technology is used to identify insects, it is necessary to convert color images into gray images, and the quality of converting color images into gray images will significantly affect the image recognition effect of insects. In this paper, four commonly used methods of converting color images into gray images are compared with the newly proposed color channel comparison method to show the advantages of the new method.

Color feature is the most widely used visual feature in image retrieval. The main reason is that color is often very related to the objects or scenes contained in the image. To extract color features, we must choose a suitable color space to calculate color features and choose a suitable method to quantify color features ¹⁰. Color features are rotation invariant and scale invariant, which are insensitive to complex background, image size and orientation, and show strong robustness.

Color moment is a very simple and effective color feature. Compared with color histogram, another advantage of this method is that there is no need to vectorize the feature. The mathematical expression of the three low-order moments of color is:

Where p_ij is the value of the i color component of the j pixel in the image, and N is the number of pixels. Therefore, the color moment of the image only needs 9 components, which is very concise compared with other color features.

Morphological characteristics are a kind of characteristics that describe objects intuitively, and it is one of the important ways to distinguish different kinds of insects in traditional insect taxonomy. Eccentricity describes the compactness of a region in a certain length, and the general method to calculate eccentricity is to find the ratio of the major axis to the minor axis of the boundary. It can be seen that eccentricity can roughly distinguish slender objects from more square objects.

Among them, W is the sum of the trunk length and twice the wing length, and L is the body length.

The image descriptor used by HOG for human target detection can well represent the gradient intensity and gradient direction distribution in the local area of the image ¹¹. Let the number of cells be λ, the number of directions in each cell be β, and the number of blocks be δ, then the final feature vector dimension of the image is:

The sample image used in this paper is adjusted to the size of 64×128, and the overlap rate of block blocks is set to 50%. The HOG feature of the gray image is calculated, and finally the HOG feature of the insect image is 108 dimensions.

Insect species identification based on texture features. Texture is an attribute that reflects the spatial distribution of pixel gray level in a region and is an important factor to describe the characteristics of objects. The run-length of a certain gray value is generally defined as the number of pixels that are continuous, collinear and have the same gray value in the detected area. At present, many scholars use gray value run-length matrix to express texture features and realize insect recognition, but the application of other texture feature operators has not been studied.

Texture features describe the spatial distribution law of domain gray level of pixels in an image. Different species of insects have different texture features, and texture feature analysis is an important research content of insect species identification. Micro-filtering calculation is carried out with a template in a 3×3 or 5×5 window, and the sum of convolution is taken as the output new image, and then each pixel value in the convolved image is replaced by the mean square error or absolute value average in a 20×20 local window centered on the pixel.

Carrying out two-dimensional Fourier transform on the image;

According to the separation characteristics of two-dimensional Fourier transform:

According to the above separation form, two-dimensional Fourier transform can be realized by applying one-dimensional Fourier transform twice.

2.2

Intelligent insect species identification

In the process of feature extraction, we always extract as many parameters as possible to describe the features of objects as completely as possible. However, from the method of feature extraction, many features are not independent and redundant. The result of feature optimization affects the performance of the classifier. A good feature optimization result set will make the classifier easy to design, while a poor feature optimization result set will make the classifier judge the classification for a long time and increase the misjudgment rate. The performance of classifier directly determines the overall performance of classification and recognition system, so the design of classifier is very important.

Transplanting content-based image retrieval into insect image recognition application can easily realize insect image recognition and comparison. However, due to the one-sidedness of a single feature extraction method, the expected effect cannot be achieved by using only one feature extraction method. For insect images, using feature fusion technology will obtain more comprehensive insect image features, which will greatly improve the recognition accuracy and improve the recognition efficiency of the system.

An important problem to be solved in comprehensive feature retrieval is the normalization of algorithm matching results. Because the similar distances obtained by different algorithms are not comparable to each other, that is, their expectations and variances may be far apart, so they must be normalized. This process can be completed by Formula (8).

Where μ_k,σ_k refers to the mean and standard deviation of the sequence v_k.

SVM has many unique advantages in solving small sample, nonlinear and high-dimensional pattern recognition, and can be extended to other machine learning problems such as function fitting. SVM is a two-class classifier, and its main idea is to find an optimal classification surface under the condition of linear separability, and the farther the two kinds of data to be classified are from the classification surface, the better. The flow chart of intelligent insect species identification based on computer vision is shown in Figure 1.

Figure 1.

Intelligent insect species identification process based on computer vision

In the learning process, a certain number of insect images are selected as training samples, and the feature vectors of insects are extracted as the input vectors of SVM training, and the input vectors are learned to get the best classification model. In the process of testing, the feature vectors of test images are extracted and input into the trained classification model, and finally the classification categories of test samples are obtained.

Color and texture are regarded as different views, which are respectively learned in the subspaces of color and texture. After that, the two classifiers are used to classify unlabeled images, and those unlabeled images classified by the two classifiers are selected to be labeled. SVM learns the labeled samples in the whole view and gets a classification surface h. The unlabeled images are divided into positive feedback and negative feedback by h, and m positive feedback pictures are fed back to users for labeling.

The SVM is trained in the whole view to obtain the classifier h, and the unlabeled set D_{Unmarked} is classified, and the top m images are labeled as :

When we encounter the problem of linear inseparability, we can consider mapping the problem to a higher dimensional space to make it linearly separable. However, mapping the features of low-dimensional space to high-dimensional space and calculating its inner product is a heavy task, even impossible ¹². In order to solve this problem, only x_i, x is the vector in the inner product, so the optimal classification surface g(x) can be expressed as:

The auxiliary non-negative variable a_i is called Langrange multiplier.

In order to solve it conveniently, the concept of kernel function is introduced. Let the form of kernel function be K(w,x), then the classifier is:

The parameter a, y, b in the above formula has not changed, that is to say, in the process of solving, whenever the inner product is required, the kernel function can be used to calculate, so that a can be combined with the kernel function to get the classifier.